A Generalizable and Accurate AI Model, Tested in Diverse Contexts

A study published in JAMA Network Open evaluates an artificial intelligence model developed by Google Research and Google for Health, designed to estimate gestational age from ultrasounds performed by non-specialized operators. The system, based on a technique called « blind sweep ultrasonography » (blind-sweep ultrasound), was trained on data from the United States (North Carolina) and Zambia, then refined with participants from Chicago (United States) and prospectively tested in the same city as well as in Nairobi (Kenya). No data from Nairobi was used for model adjustment, confirming its robustness in different clinical environments.

The results show that the system’s mean absolute error (4.18 days) is comparable to that of expert sonographers (4.48 days), according to data reported by Google Research. This performance was validated even in complex cases, such as multiple pregnancies, and the model easily adapts to new equipment, like the Clarius C3 probe, with minimal adjustments.

Gestational age serves as the reference for all screenings, treatments, and decisions regarding delivery dates, a precision that could transform prenatal care in under-equipped areas.

A Solution to Address Limited Access to Prenatal Ultrasounds

Two-thirds of the global population lack access to prenatal ultrasounds due to logistical, financial, or human constraints, notes Google Research. The proposed AI model aims to reduce this inequality by enabling non-trained operators to obtain reliable gestational age estimates, a crucial parameter for medical monitoring. Nichole Young-Lin, an obstetrician cited in the study, emphasizes that « gestational age serves as the reference for all screenings, treatments, and decisions regarding delivery dates. » Improved accuracy in these assessments could therefore enhance patient care, particularly in resource-limited areas.

What This Means Here: Potential Applications for Benin and West Africa

In Benin, as in much of West Africa, health systems face similar challenges: a shortage of specialized medical staff, inequalities in access to advanced medical equipment, and regional disparities in the quality of prenatal care. A model like the one developed by Google Research could be adapted to reduce these gaps by enabling clinics and hospitals, even those in rural areas, to use simplified ultrasounds with AI assistance. This could also ease the burden on healthcare professionals by automating part of routine diagnostics, while improving the reliability of data collected for public health programs.

However, its deployment would depend on several factors: acceptance by local health authorities, the availability of compatible equipment, and the training of end users. Collaboration with actors like the World Health Organization (WHO), which supported the study, could facilitate its integration into regional health policies.

Technology designed to evolve with local constraints

One of the key strengths of this model lies in its ability to adapt to different types of ultrasound equipment, as demonstrated by tests conducted with probes from various brands (Sonosite, GE, Butterfly Networks, Clarius C3). This flexibility could be an advantage for countries where medical equipment is often mismatched or outdated. Additionally, the 'blind sweep' method does not require advanced technical skills to perform examinations, which could accelerate its adoption in contexts where human resources are scarce.

Google Research also notes that the model was designed to be deployed in resource-limited environments, an essential criterion for its adoption in Sub-Saharan Africa. The next steps could include additional tests in other regions of the continent, in collaboration with local partners, to refine its effectiveness and operational utility.